Predicting driver injury severity and identifying factors influencing crash outcomes are crucial for developing effective traffic safety measures. This study focuses on estimating driver injury severity (uninjured, injured, or killed) and determining critical factors affecting crash outcomes. A hybrid framework combining Deep Neural Networks (DNNs) and Random Forest (RF) is proposed, where a DNN extracts features and RF performs the final classification, leveraging ensemble methods. The results were compared with those of well-known methods (e.g., kNN, XGBoost), with the hybrid approach achieving the best performance (0.92 accuracy, 0.89 F1-macro, 0.91 F1-micro scores) in predicting injury severity. The results showed that crash type, vehicle type, driver fault, intersection type, season, time, and road type had the greatest impact, while factors like pavement condition and driver gender had minimal influence. To the best of our knowledge, this is the first study to combine DNN-based feature extraction with RF classification in the context of traffic injury severity prediction. The framework offers a new approach for drivers and policymakers, providing a deeper understanding of driver injury severity prediction and its underlying factors.

Predicting driver injury severity is critical for enhancing road safety, but it is complicated because fatal accidents inherently create class imbalance within datasets. This study conducts a comparative analysis of machine-learning (ML) and deep-learning (DL) models for multi-class driver injury severity prediction using a comprehensive dataset of 107,195 traffic accidents from the Adana, Mersin, and Antalya provinces in Turkey (2018–2023). To address the significant imbalance between fatal, injury, and non-injury classes, the hybrid SMOTE-ENN algorithm was employed for data balancing. Subsequently, feature selection techniques, including Relief-F, Extra Trees, and Recursive Feature Elimination (RFE), were utilized to identify the most influential predictors. Various ML models (K-Nearest Neighbors (KNN), XGBoost, Random Forest) and DL architectures (Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Recurrent Neural Network (RNN)) were developed and rigorously evaluated. The findings demonstrate that traditional ML models, particularly KNN (0.95 accuracy, 0.95 F1-macro) and XGBoost (0.92 accuracy, 0.92 F1-macro), significantly outperformed DL models. The SMOTE-ENN technique proved effective in managing class imbalance, and RFE identified a critical 25-feature subset including driver fault, speed limit, and road conditions. This research highlights the efficacy of well-preprocessed ML approaches for tabular crash data, offering valuable insights for developing robust predictive tools to improve traffic safety outcomes.

The K Nearest Neighbor (KNN) method is one of the most used machine-learning algorithms due to its simplicity. However, as the dataset grows, the running speed of the KNN algorithm decreases considerably. In this study, the performance of the two parallel KNN algorithms (i.e. GPU-based TUKNN and GPU-SME-KNN) were compared with the sequential KNN algorithm using five different datasets which have 520, 5110, 78,096, 500,000, and 2,916,697 samples. According to the experiments performed on these datasets, it is observed that the TUKNN and GPU-SME-KNN algorithms are 10.4 times and 347.3 times faster than the sequential KNN algorithm on average, respectively. TUKNN's average CPU consumption is 7.93%, and average RAM consumption is 1.49%. The average CPU consumption of the GPU-SME-KNN algorithm provides an efficiency of 8.59%, while the average RAM consumption increases by 6.29%.

Distance-based machine learning algorithms are widely used due to their simplicity and flexibility. However, they often suffer from sensitivity to noise, high computational cost, and instability in high-dimensional and complex data spaces. To address these limitations, this study introduces a novel geometric learning framework based on Bipolar Metric Spaces (BMSs). The proposed framework extends classical metric structures by incorporating a quadrilateral inequality constraint, enabling the modeling of both similarity and contrast relationships across data points. Within this framework, geometric regularization is achieved through a structural filter mechanism that filters out structurally inconsistent assignments and stabilizes decision boundaries. Importantly, the proposed approach is not limited to specific algorithms and can be integrated into a broad class of distance-based learning methods. To demonstrate its effectiveness, the framework is instantiated on K-Nearest Neighbor (KNN) and K-means algorithms, resulting in Bipolar K-Nearest Neighbor (BP-KNN) and Bipolar K-means (BP-K-means). Theoretical analysis shows that the proposed framework improves Lipschitz stability, reduces variance, and enhances generalization performance. The models were tested on six different datasets, including benchmark datasets (Iris, Wine, and Breast Cancer) and real-world datasets (Abalone, CDC Diabetes, and Malicious Website Features (MWS)). Theoretical complexity analysis, supported by Floating Point Operation (FLOP) quantification, reveals that the bipolar approach offers a more stable decision boundary and significantly faster convergence, corroborating the observed efficiency improvements. Notably, on the large-scale MWS dataset, BP-KNN achieved an 11.6-fold increase in inference speed and a 75% reduction in RAM usage compared with the standard KNN. These results demonstrate that BMS provides a unified and scalable geometric regularization framework for distance-based learning, offering a principled solution to the trade-off between accuracy, robustness, and computational efficiency.